Driver’s Stress Detection using Skin Potential Response Signals

Stress in car drivers represents a risk, especially for professional car drivers which are more likely to be exposed to it for prolonged periods. A persisting stress state leads to mental and physical pathologies and increases the probability of causing accidents. Thus, the monitoring of drivers’ mental state could allow an immediate action before the problem degenerates. In the present work, two main methods to detect a subject’s sympathetic reaction to stress are developed. Method 1: we measure Skin Potential Response (SPR) and record the Steering Wheel angle excursion. Then, we process the measured signals with adaptive filters which remove the component related to Motion Artifact, exploiting the relation between hand movements to handle the Steering Wheel and MA. Next, we process the obtained Stress signal with a Smooth Nonlinear Energy Operator (SNEO) to locate stress events. Experiments which allow to define a ground-truth for stress events recognition, show that, by appropriately processing, it is possible to efficiently detect stress events, obtaining a mean Recall of 95 %. Method 2: A double channel SPR sensor is used to measure SPR from both the hands, and Electrocardiogram is recorded with a triple channel ECG sensor. SPR measurements are processed through an algorithm which selects the smoother signal. In this way we obtain a Stress signal without the Motion Artifact component. Several experiments carried out in laboratory and with a real car professional simulators, reveal the efficacy of the proposed system, which outperformed Principal Component Analysis and Independent Component Analysis. Next, machine learning techniques are employed to classify features obtained from the Stress signal and from Heart Rate Variability. In particular, we considered a Support Vector Machine (SVM) and a feed-forward Neural Network (NN). The system is tested in experiments carried out on a professional driving simulator. Classifying 15 seconds long time intervals, we obtained a balanced accuracy of 76.72% for SVM and 77.15% for NN. Applying a relabeling method based on the previous time intervals, the performances raised to 78.74% for SVM and 78.26% for NN. We then, tested the ability to identify stress intervals and obtained a balanced accuracy of 89.58% for SVM and 91.92% for NN.